cmd/internal/obj/arm64: avoid unnecessary literal pool usage for moves

In a number of load and store cases, the use of the literal pool can be
entirely avoided by simply adding or subtracting the offset from the
register. This uses the same number of instructions, while avoiding a
load from memory, along with the need for the value to be in the literal
pool. Overall this reduces the size of binaries slightly and should have
lower overhead.

Updates #59615

Change-Id: I9cb6a403dc71e34a46af913f5db87dbf52f8688c
Reviewed-on: https://go-review.googlesource.com/c/go/+/512539
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: Joel Sing <joel@sing.id.au>

2 files changed, 74 insertions, 5 deletions

diff --git a/src/cmd/asm/internal/asm/testdata/arm64.s b/src/cmd/asm/internal/asm/testdata/arm64.s
index 54f4de76d8..1aa74caa26 100644
--- a/src/cmd/asm/internal/asm/testdata/arm64.s
+++ b/src/cmd/asm/internal/asm/testdata/arm64.s
@@ -557,7 +557,38 @@ TEXT	foo(SB), DUPOK|NOSPLIT, $-8
 	FMOVQ	65520(R10), F10 // 4afdff3d
 	FMOVQ	64(RSP), F11    // eb13c03d
 
-// large aligned offset, use two instructions(add+ldr/store).
+// medium offsets that either fit a single instruction or can use add+ldr/str
+	MOVD -4095(R17), R3                        // 3bfe3fd1630340f9
+	MOVD -391(R17), R3                         // 3b1e06d1630340f9
+	MOVD -257(R17), R3                         // 3b0604d1630340f9
+	MOVD -256(R17), R3                         // 230250f8
+	MOVD 255(R17), R3                          // 23f24ff8
+	MOVD 256(R17), R3                          // 238240f9
+	MOVD 257(R17), R3                          // 3b060491630340f9
+	MOVD 391(R17), R3                          // 3b1e0691630340f9
+	MOVD 4095(R17), R3                         // 3bfe3f91630340f9
+
+	MOVD R0, -4095(R17)                        // 3bfe3fd1600300f9
+	MOVD R0, -391(R17)                         // 3b1e06d1600300f9
+	MOVD R0, -257(R17)                         // 3b0604d1600300f9
+	MOVD R0, -256(R17)                         // 200210f8
+	MOVD R0, 255(R17)                          // 20f20ff8
+	MOVD R0, 256(R17)                          // 208200f9
+	MOVD R0, 257(R17)                          // 3b060491600300f9
+	MOVD R0, 391(R17)                          // 3b1e0691600300f9
+	MOVD R0, 4095(R17)                         // 3bfe3f91600300f9
+	MOVD R0, 4096(R17)                         // 200208f9
+	MOVD R3, -4095(R17)                        // 3bfe3fd1630300f9
+	MOVD R3, -391(R17)                         // 3b1e06d1630300f9
+	MOVD R3, -257(R17)                         // 3b0604d1630300f9
+	MOVD R3, -256(R17)                         // 230210f8
+	MOVD R3, 255(R17)                          // 23f20ff8
+	MOVD R3, 256(R17)                          // 238200f9
+	MOVD R3, 257(R17)                          // 3b060491630300f9
+	MOVD R3, 391(R17)                          // 3b1e0691630300f9
+	MOVD R3, 4095(R17)                         // 3bfe3f91630300f9
+
+// large aligned offset, use two instructions(add+ldr/str).
 	MOVB	R1, 0x1001(R2) 		// MOVB		R1, 4097(R2)		// 5b04409161070039
 	MOVB	R1, 0xffffff(R2)	// MOVB		R1, 16777215(R2)	// 5bfc7f9161ff3f39
 	MOVH	R1, 0x2002(R2)		// MOVH		R1, 8194(R2)		// 5b08409161070079
diff --git a/src/cmd/internal/obj/arm64/asm7.go b/src/cmd/internal/obj/arm64/asm7.go
index 72e6aadce4..05cf62773e 100644
--- a/src/cmd/internal/obj/arm64/asm7.go
+++ b/src/cmd/internal/obj/arm64/asm7.go
@@ -1959,6 +1959,10 @@ func (c *ctxt7) loadStoreClass(p *obj.Prog, lsc int, v int64) int {
 	}
 
 	needsPool := true
+	if v >= -4095 && v <= 4095 {
+		needsPool = false
+	}
+
 	switch p.As {
 	case AMOVB, AMOVBU:
 		if cmp(C_UAUTO4K, lsc) || cmp(C_UOREG4K, lsc) {
@@ -4015,10 +4019,13 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		o1 |= uint32(p.From.Reg&31)<<5 | uint32(p.To.Reg&31)
 
 	case 30: /* movT R,L(R) -> strT */
-		// if offset L can be split into hi+lo, and both fit into instructions, do
+		// If offset L fits in a 12 bit unsigned immediate:
+		//	add $L, R, Rtmp  or  sub $L, R, Rtmp
+		//	str R, (Rtmp)
+		// Otherwise, if offset L can be split into hi+lo, and both fit into instructions:
 		//	add $hi, R, Rtmp
 		//	str R, lo(Rtmp)
-		// otherwise, use constant pool
+		// Otherwise, use constant pool:
 		//	mov $L, Rtmp (from constant pool)
 		//	str R, (R+Rtmp)
 		s := movesize(o.as)
@@ -4032,6 +4039,20 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		}
 
 		v := c.regoff(&p.To)
+		if v >= -256 && v <= 256 {
+			c.ctxt.Diag("%v: bad type for offset %d (should be 9 bit signed immediate store)", p, v)
+		}
+		if v >= 0 && v <= 4095 && v&((1<<int32(s))-1) == 0 {
+			c.ctxt.Diag("%v: bad type for offset %d (should be 12 bit unsigned immediate store)", p, v)
+		}
+
+		// Handle smaller unaligned and negative offsets via addition or subtraction.
+		if v >= -4095 && v <= 4095 {
+			o1 = c.oaddi12(p, v, REGTMP, int16(r))
+			o2 = c.olsr12u(p, c.opstr(p, p.As), 0, REGTMP, p.From.Reg)
+			break
+		}
+
 		hi, lo, err := splitImm24uScaled(v, s)
 		if err != nil {
 			goto storeusepool
@@ -4054,10 +4075,13 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		o2 = c.olsxrr(p, int32(c.opstrr(p, p.As, false)), int(p.From.Reg), int(r), REGTMP)
 
 	case 31: /* movT L(R), R -> ldrT */
-		// if offset L can be split into hi+lo, and both fit into instructions, do
+		// If offset L fits in a 12 bit unsigned immediate:
+		//	add $L, R, Rtmp  or  sub $L, R, Rtmp
+		//	ldr R, (Rtmp)
+		// Otherwise, if offset L can be split into hi+lo, and both fit into instructions:
 		//	add $hi, R, Rtmp
 		//	ldr lo(Rtmp), R
-		// otherwise, use constant pool
+		// Otherwise, use constant pool:
 		//	mov $L, Rtmp (from constant pool)
 		//	ldr (R+Rtmp), R
 		s := movesize(o.as)
@@ -4071,6 +4095,20 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		}
 
 		v := c.regoff(&p.From)
+		if v >= -256 && v <= 256 {
+			c.ctxt.Diag("%v: bad type for offset %d (should be 9 bit signed immediate load)", p, v)
+		}
+		if v >= 0 && v <= 4095 && v&((1<<int32(s))-1) == 0 {
+			c.ctxt.Diag("%v: bad type for offset %d (should be 12 bit unsigned immediate load)", p, v)
+		}
+
+		// Handle smaller unaligned and negative offsets via addition or subtraction.
+		if v >= -4095 && v <= 4095 {
+			o1 = c.oaddi12(p, v, REGTMP, int16(r))
+			o2 = c.olsr12u(p, c.opldr(p, p.As), 0, REGTMP, p.To.Reg)
+			break
+		}
+
 		hi, lo, err := splitImm24uScaled(v, s)
 		if err != nil {
 			goto loadusepool